Adversarially robust learning for security-constrained optimal power flow